Open Weights Just Got Bought

Two acquisitions in six weeks settled a question the AI industry has argued about since 2023: are open-weight models a research norm or strategic infrastructure? The answer arrived in the form of purchase prices. On 3 September NVIDIA agreed to buy Hugging Face, the platform that hosts most of the world's open models, for $12.93 billion. Three weeks earlier, Stripe agreed to buy OpenRouter, the marketplace through which developers route requests to more than 400 models. Infrastructure is what companies pay billions to own.

The hub

Hugging Face hosts more than 3 million models, 500,000 datasets and 1 million applications, used by more than 18 million developers and over 200,000 companies, according to NVIDIA's announcement. Reuters reported NVIDIA will pay about $11.9 billion to investors and offer up to $1 billion in equity retention to staff, with the deal expected to close in the first half of 2027 pending regulatory approval. NVIDIA is already the largest contributor of open models and data to the platform, with more than 500 models and 250 datasets released.

Jensen Huang's pledge was explicit: Hugging Face will remain an open platform for the entire ecosystem, developers will choose their models, frameworks, clouds and compute, and NVIDIA compute will not be required to build or deploy through it. Hugging Face chief executive Clément Delangue framed the sale as open-source AI needing 'more compute, more support, more collaboration and more visibility' to compete with closed APIs at scale. The strategic logic for NVIDIA is straightforward: an open ecosystem creates the pipeline of developers and workloads that eventually runs on its chips, whether or not the platform requires them.

The switch

OpenRouter's own announcement on 19 August described what Stripe is buying: the first and largest model marketplace and gateway, processing more than 10 trillion tokens a day from 400-plus models for over 10 million developers and companies, with inference volume growing at least tenfold every year since founding. Bloomberg reported the price at more than $7 billion, a steep premium on the $1.3 billion valuation of OpenRouter's Series B three months earlier. A CNBC analysis in July found that Chinese-origin models accounted for 46% of US enterprise token usage on the platform.

That last figure is why the routing layer is strategic. Whoever owns the switch sees, in real time, which models enterprises actually use and what they pay. Stripe has said routing decisions will stay driven by what is best for the user. It has also, in effect, bought the market data for the intelligence economy.

What got released

The models being routed and hosted got bigger and more permissive at the same time. Moonshot AI launched Kimi K3 via API on 16 July and published its weights on Hugging Face at the end of the month: 2.8 trillion total parameters in a mixture-of-experts design with 896 experts, 104 billion parameters active per token according to its model card, a native vision encoder and a context window of 1,048,576 tokens. It is the largest open-weight model released to date. Its download is about 594 GB; serving it takes cluster-class hardware.

DeepSeek's V4 family, previewed in April and generally available by August, ships under the MIT licence: V4-Pro at 1.6 trillion total parameters with 49 billion active, and V4-Flash at 284 billion total with about 13 billion active, both with 1-million-token context. Alibaba released Qwen3.8-27B on 14 August under Apache 2.0, a dense multimodal model that runs on a single GPU, and two days earlier published weights for its 2.4-trillion-parameter Qwen3.8-Max under a custom licence rather than Apache. That split is the pattern of the moment: permissive licences for the models you can run, bespoke terms for the flagships.

The gap between open and usable

The frontier of open weights now requires a data centre. Kimi K3 and Qwen3.8-Max are open in the sense that anyone may download them, and closed in the sense that almost nobody can run them. The practical open frontier for most enterprises is the 20- to 30-billion-parameter tier: Qwen3.8-27B at roughly 17 GB quantised, DeepSeek V4-Flash for those with a few GPUs. That is where sovereignty, data residency and cost arguments for open weights are actually cashed in, and it is where the acquirers' platforms matter most, because discovery and routing decide which of these models enterprises adopt.

Who benefits, who is at risk

Beneficiaries: NVIDIA and Stripe, which now own the two front doors to open models; Chinese labs whose releases dominate open-weight usage; and enterprises that get cheaper, more capable self-hostable models every quarter. At risk: closed-API vendors whose pricing depends on a capability gap that keeps narrowing, and developers who read 'open' on a licence without reading the terms beneath it.